Using Artificial Intelligence to Predict Implantable Collamer Lens Vault: A Low Parameter-Dependent Model for Better Surgical Outcomes.

Abstract

Purpose: The purpose of this study was to predict the vault of implantable collamer lens using artificial intelligence (AI) and interpret the contributions of each parameter.

Methods: Extreme Gradient Boosting (XGBoost), a machine learning algorithm, was applied to construct a vault prediction model. The dataset included 247 eyes from Peking University Third Hospital, split into training and test sets (4:1), plus 50 eyes from Beau Care Clinic for external validation. The model was trained and tested by samples with missing and anomalous values to enhance its robustness. Model performance was assessed using mean absolute error (MAE), root mean square error (RMSE), and median absolute error (MedAE). SHapley Additive exPlanations (SHAP) was used to interpret the model's predictions.

Results: We found weak linear correlation between preoperative parameters and vaults (all |r| ≤ 0.30). Therefore, a nonlinear model was constructed. It achieved the following performance on the test set: MAE = 117.85 µm, RMSE = 146.92 µm, and MedAE = 108.94 µm. On the external validation set, corresponding metrics were 130.99 µm, 154.24 µm, and 116.51 µm, respectively. SHAP revealed horizontal sulcus-to-sulcus distance (STS), horizontal compression (HC), anterior chamber depth (ACD), and white-to-white distance (WTW) had positive influences on the vault, whereas lens thickness (LT) and crystalline lens rise (CLR) had negative effects. Female subjects also tended to have higher vaults.

Conclusions: A low parameter-dependent implantable collamer lens (ICL) vault prediction model which exhibits great robustness was constructed.

Translational relevance: The use of AI to predict the vault after ICL implantation can reduce the abnormal postoperative vault and improve the safety of ICL implantation.